Well, it kinda breaks the game, doesn't it?

Can you decouple rendering from simulation? I guess so, since it's easy to test whether a point lies below or above the curve. If at least what was simulated were reliable that level would be (somewhat) playable.

I would suggest having reliable visual sampling too (samples in absolute space, growing in number with zoom out), but I'm not aware of the actual performance implications.

I still think there's something else going on besides sampling error. Why is it (visually) working correctly for x*7 until I sled? Why does it change so dramatically over time?